The invention concerns a process for reducing noise signals in telecommunications (TC) systems for the transmission of acoustic useful signals, in particular human speech.
A known process for noise reduction is so-called xe2x80x9cspectral subtractionxe2x80x9d, that is described in the publication xe2x80x9cA new approach to noise reduction based on auditory masking effectsxe2x80x9d by S. Gustafsson and P. Jax, ITG Conference, Dresden, 1998, for example. This involves a spectral noise reduction method in which an acoustic masking threshold (for example following the MPEG standard) is taken into consideration.
During natural communication between humans, the amplitude of the spoken language is usually adapted to the acoustic environment automatically. In the case of speech communication between distant locations, however, the interlocutors are not in the same acoustic surroundings and each is not therefore aware of the acoustic situation at the location of the other interlocutor. The problem therefore gets worse if, because of his/her acoustic environment, one of the parties is forced to speak very loudly while the other party in a quiet acoustic environment produces voice signals with lower amplitude.
Noise problems are particularly acute in new communication systems applications, for example mobile telephones, in which the terminals are made so small that a direct spatial juxtaposition between loudspeaker and microphone cannot be avoided. Because of the direct sound transmission, in particular structure-borne noise between loudspeaker and microphone the acoustic interference signal can have the same order of magnitude as the useful signal of the speaker at the respective terminal or its amplitude can even exceed this signal. Such a noise problem also occurs to a significant degree in the case of several terminals arranged spatially adjacent to each other, for example in an office or conference room with a number of telephone connections, since a coupling takes place from each loudspeaker signal to each microphone.
Added to this is the problem that on a telecommunications channel xe2x80x9celectronically generatedxe2x80x9d noise also occurs and is transmitted as background along with the useful signal. In order to increase comfort while making a telephone call, one therefore endeavours to keep each type of noise as low as possible in comparison to the useful signal.
Finally, one also endeavours to reduce or completely suppress interference signals such as undesirable background noise (traffic noise, factory noise, office noise, canteen noise, aircraft noise, etc.).
In the known compander process, such as described in DE 42 29 912 A1, the degree of noise reduction is determined by a fixed, predetermined transfer function. First of all, the compander has the property of transmitting voice signals at a specific (previously set) xe2x80x9cnormal speech signal levelxe2x80x9d (sometimes referred to as normal loudness) virtually unchanged from its input to the output. If, however, the input signal now becomes too loud, for example because a speaker comes too close to its microphone, then a dynamic compressor limits the output level to virtually the same value as in the normal case, by reducing the actual gain in the compander linearly with increasing input loudness. Due to this characteristic the speech at the output of the compander system remains more or less at the same loudnessxe2x80x94irrespective of how widely the input loudness fluctuates. On the other hand, if a signal with a level that is lower than the normal level is now applied to the input of the compander, then the signal is additionally attenuated by reducing the gain in order to transmit background noise that is attenuated as far as possible. The compander thus consists of two partial functions, a compressor for the speech signal levels that are higher than or equal to a normal level, and an expander for signal levels that are lower than the normal level.
In the case of the above-mentioned spectral subtraction, to this end the noise is first measured in the speech pauses and continuously stored in a memory in the form of a power spectral density. The power spectral density is obtained via a Fourier transformation. When speech occurs, the stored noise spectrum is subtracted from the current disturbed speech spectrum xe2x80x9cas best current estimated valuexe2x80x9d, then transformed back into the time domain in order by this means to obtain a noise reduction for the disturbed signal.
A disadvantage of such methods is the complicated determination of this acoustic masking threshold and the execution of all computing operations associated with this method. A further disadvantage of spectral subtraction is that due to the process of a basically inaccurate spectral noise estimate and subsequent subtraction, errors which are perceptible as xe2x80x9cmusical tonesxe2x80x9d, also occur in the output signal.
With extended spectral signal processing, which is also described in the citation mentioned at the beginning, the power spectral densities are estimated for the noise and for the speech itself with the aid of a spectral subtraction. Knowing these partial spectra, a spectral acoustic masking threshold RT(f) is then calculated for the human ear with the aid of MPEG Standard rules, for example. Using this masking threshold and the estimated spectra for noise and speech, and following a simple rule, a filter passband curve H(f) is calculated, which is configured so that essential spectral components of the speech are transmitted with as little modification as possible and spectral components of the noise are reduced as much as possible.
The original disturbed speech signal is then passed only through this filter to obtain a noise reduction for the disturbed signal by these means. The advantage of this method is now that xe2x80x9cnothing is added to or subtracted fromxe2x80x9d the disturbed signal and therefore errors in the estimations are less perceptible or even scarcely perceptible. A disadvantage is again the considerably greater computing power.
A particular disadvantage of all these known methods is the fact that the incoming original signal undergoes a signal processing process prior to the actual subtraction of a noise signal that is always simulated, and is therefore basically corrupted.
In contrast, the object of the present invention is to present a process with least possible complexity having the features described at the outset, in which a noise reduction or noise suppression is achieved in an uncomplicated technical manner, and in which the original signal remains uncorrupted right up to the actual noise subtraction. At the same time, with simple means, in particular with less computing power than previously, the process should enable an overall acoustic impression to be produced, which is as agreeable as possible to the human ear and which, according to taste, can be matched to individual requirements. Finally, the new process should be capable of being implemented completely independently of the speech signal processing requirements and thus enable simple optimisation to the spectral processing requirements of noise signals.
This object is achieved according to the invention in both a simple and effective manner by the following process steps:
(a) Determining by means of speech pause detection when a speech pause is contained in the mixture of useful signals and interference signals to be transmitted, or when a speech pause is present;
(b) Branching the incoming TC signal from the main signal path and using a Fourier transformation on the branched TC signal to generate a frequency spectrum of the branched TC signal;
(c) Storing in a buffer memory the last frequency spectrum recorded during the last speech pause;
(d) Using an inverse Fourier transformation on the last respective recorded frequency spectrum to generate a simulated noise signal;
(e) Subtracting the simulated noise signal in the time domain from the current incoming TC signal.
Due to the separate simulation of the noise signal in the frequency domain independently of a processing of the original speech signal, the process according to the invention allows a direct subtraction of the simulated noise signal from the original, uncorrupted input signal, which undergoes neither a Fourier transformation nor an inverse Fourier transformation. With suitable phase correction in the frequency domain, noise subtraction from the original signal is even possible without a time delay. At the same time the process according to the invention is less complex than the above-mentioned known processes from the prior art, requires less computing power and results in a better frequency resolution.
By separating the noise simulation from the transmission of the original signal, the process according to the invention enables, in a particularly preferred variant, in step (d), only a selected part of the generated frequency spectrum to be utilised for the generation of the simulated noise signal. The computing power required for implementing the process according to the invention is thus further minimised or the process itself can be carried out more rapidly.
A development of this process variant is characterised by the fact that the selection of the part of the frequency spectrum used for the generation of the simulated noise signal is made in accordance with psycho-acoustic criteria implementing the mean values of the perception spectrum of the human ear.
In this case the value for the noise signal to be simulated is determined not only from the instantaneous power value of an original signal in speech pauses alone, but also from a weighted spectral characteristic of the corresponding signal and overall, via the function obtained in this way, achieves an acoustically correct noise reduction, that is to say one that is psycho-acoustically pleasant-sounding.
Since there is no measure for an acoustically pleasant-sounding noise reduction, that can be easily represented, all quality assessments rely on extensive listening tests which are then evaluated by means of statistical methods optimised for this purpose, in order to obtain a weighting rule (similar to speech codecs).
The basic procedures for this are to be found in the text book xe2x80x9cPsychoacousticsxe2x80x9d by E. Zwicker, Springer-Verlag Berlin, 1982, in particular pages 51 to 53, for example.
Due to the psycho-acoustic evaluation, not only can the perceptible quality of the overall signal be optimised, but further savings in the necessary computing power are possible if, for example, masking effects are utilised or only those frequencies that are clearly caused by sources of noise or interference are taken into consideration.
In an alternative development of the above process variant, the selection of the part of the frequency spectrum used for the generation of the simulated noise signal is made in such a way that only discrete frequencies of the spectrum are considered, and that the spacing between the discrete frequencies is made to steadily increase towards the higher frequencies and preferably in accordance with a logarithmic function. The frequency resolution is thus better matched to the perception of the human ear.
These developments can be further improved by dividing the selected part of the frequency spectrum into previously determined frequency groups, and selecting in each frequency group only the frequency or frequency band, respectively, that has the highest signal energy within the frequency group and further utilising this for the generation of the simulated noise signal. This selection achieves a large reduction in the frequencies to be computed for constant audible or perceptible quality, which results in the computing power for the process being further reduced and the quality of the output signal being further enhanced.
If is particularly advantageous if the selection of the frequency or the frequency band, respectively, having the highest signal energy within the frequency group is made prior to step (c) or step (d), respectively. By selecting a specific frequency from a frequency group, differences in the signal energy can be detected very easily.
A process variant in which in step (b) the frequency spectrum of the branched TC signal is generated only in a predetermined frequency band, is also advantageous. Provided the interference source has only a limited frequency spectrum, again considerable computing power can be saved with this measure. For example in powered vehicles, interference sources having a frequency band of up to a maximum of 1 only KHz are considered since the interference signal is in the main formed by low-frequency sound generation (engine, gearbox, motion noise, etc.).
A particularly simple process variant is characterised by the fact that in step (b) and/or step (d) a discrete Fourier transformation or an inverse discrete Fourier transformation is used, where time-discrete amplitude values are sampled from the incoming TC signal at a sampling frequency fT.
In a preferred development of the process variant, a fast Fourier transformation (FFT) is utilised in step (b). If a wide frequency range together with high frequency resolution are to be covered, this procedure allows analysis with lowest computing power. The FFT is then particularly useful if more than 128 frequency lines have to be computed, for example.
Advantageously, an inverse discrete Fourier transformation (IDFT) can be employed in step (d). This allows a signal synthesis to be implemented with lowest computing power if a selected spectrum is processed, since the disadvantage of an equidistant frequency distribution in the FFT is avoided. The IDFT can therefore be advantageously utilised for a specified frequency band. The frequencies can be distributed individually. A saving in computing power with respect to the FFT is possible from a frequency resolution of less than 128 frequency lines.
In the application, savings in the computing power or quality improvements can be achieved if an inverse fast Fourier transformation (IFFT) is employed in step (d). In combination with an FFT in step (b) broadband noise sources can be processed in a particularly economical manner.
An alternative to the last-named process variant is an embodiment in which only the part of the generated frequency spectrum that lies below the half sampling frequency fT/2 is selected. Savings can thus again be made in computing power, but also in memory space utilisation.
Particularly advantageous is a variant of the process according to the invention in which a frequency spectrum that is obtained by averaging the current frequency spectrum generated in step (b) and the previously generated frequency spectra, is temporarily stored in step (c). Due to averaging, spectral lines with higher energy are found and random values or sporadic errors are systematically suppressed.
At the same time, it is particularly favourable if the averaging is carried out with different relative weighting of the currently generated frequency spectrum in different frequency bands. The natural transient response of noise sources can generally be taken into account with such differing directions. For example, the speed of an engine in a powered vehicle cannot usually be suddenly changed. Low-frequency noise sources have a higher transient recovery time than high-frequency ones. In this case the proposed weighting helps to make the adaptivity of a system stable and fast.
Here again it is particularly advantageous if the weighting is realised in accordance with psycho-acoustic criteria implementing the mean values of the perception spectrum of the human ear. As already discussed above, with psycho-acoustic weighting, the frequency-dependent transient times are matched to the auditory sensation of the human ear. An optimisation of the system with regard to naturalness, stability and adaptation time is achieved in this way.
To avoid over-compensation in the treatment of noise, in a particularly preferred variant of the process according to the invention a simulated noise signal weighted with a weighting factor a less than 1 in accordance with predetermined criteria is subtracted from the current incoming TC signal in step (e).
In an advantageous development, the weighting factor a is made a constant value that is dependent on errors in the TC system. This enables the process according to the invention to be optimised to the errors in the respective TC system in a cost-effective and simple manner. If the errors are automatically detected, then the weighting can also take place during operation.
Alternatively, the weighting factor a can be made an adjustable value in accordance with a quality scale which can be selected by the user of the TC system. Such a user-defined weighting factor allows individual, user-defined adaptation of the process according to the invention to the individual requirements. If the system according to the invention is integrated in an existing higher-order concept, a statistical value provided by the user, for example the error probability or detection rate, can be used to control the weighting factor. In the case of applications in powered vehicles, the weighting factor can also be derived from the rotational speed or linear velocity, for example.
This can be further improved by adaptively matching the weighting factor a to the current incoming TC signal. Adaptive weighting allows automatic optimisation of the noise reduction during operation. The weighting factor can be derived from statistical values such as error probability, mean value, changes of state etc. Adaptive weighting allows particularly simple and rapid adjustments to be made to the process according to the invention to suit individual conditions in the acoustic environment of the TC terminal.
A further advantageous variant of the process according to the invention is characterised by the fact that prior to step (e) a synthetic noise signal is mixed with the simulated noise signal generated in step (d). The mixing of an artificial noise signal with constant power density can be used for masking dynamic, non-stationary noise sources in the output signal.
A further variant of the process according to the invention is designed so that prior to step (e) the current incoming TC signal undergoes a specified time delay that is preferably designed so that the phase of the incoming TC signal coincides with the phase of the simulated noise signal prior to subtraction.
Provision is made in an alternative process variant to the effect that the current incoming TC signal is fed for immediate subtraction in step (e) and that prior to step (e) the phase of the simulated noise signal is matched to the phase of the current incoming TC signal. If the phase of the reproduced noise signal in the frequency domain is corrected prior to inverse transformation, the subtraction from the non-delayed signal can take place in the time domain. Disturbing signal delays can therefore be eliminated. These are unavoidable in all processes in which the useful signal (speech) takes the roundabout route via two transformations, as for example in the known spectral subtraction discussed above.
A variant of the process according to the invention in which, in addition to the detection and reduction of noise signals, the presence of echo signals is detected and/or foreseen and the echo signals suppressed or reduced, is particularly preferred. Additional echo suppression is of course only possible when the received original signal from the remote TC subscriber is included in the echo computation. This means that the noise reduction also includes echo generation that is associated with an incoming signal from the remote TC subscriber.
This process variant can be improved by dealing with the control of the reduction of noise signals separately from the reduction of echo signals.
It is also advantageous if during the period of echo reduction a synthetic noise signal is also added to the useful signal, as already discussed in detail above, in order to avoid the subjective impression of a xe2x80x9cdead linexe2x80x9d.
In particular, the synthetic noise signal can include a psycho-acoustic signal sequence (comfort noise) that is perceived as acoustically agreeable.
Alternatively, the synthetic noise signal can include a noise signal previously recorded during the current TC link, which allows a particularly xe2x80x9ctrue-to-lifexe2x80x9d current acoustical environment to be simulated.
The context of the present invention also includes a server unit, a processor module and a gate-array module supporting the process according to the invention as described above, as well as a computer program for implementing the process. The process can be realised as a hardware circuit as well as in the form of a computer program. At the present time, software programming for high-performance DSPs is preferred, since new know-how and auxiliary functions are easier to implement by modifying the software to existing basic hardware. However, processes can also be implemented as hardware modules, for example in TC terminals or telephone installations.
Further advantages of the invention are revealed in the description and the drawing. The above mentioned features and others to be mentioned later according to the invention can equally be utilised individually or jointly in any combinations. The illustrated and described embodiments are not to be construed as a final list, but rather as having an exemplary nature for the portrayal of the invention.